DyMamba: dynamic Mamba for microscopy image semantic segmentation

Segmentation of cell bodies and organelles in microscopy images is critical for biological research, particularly in scenarios with multiple regions of interest where spatial continuity is essential. The Mamba architecture, derived from State Space Models (SSMs), has recently gained attention for efficiently modeling long-range dependencies in sequences, achieving excellent results in both natural and medical image segmentation. However, in vision tasks, current Mamba scanning strategies mainly focus on raster-scanning and local-scanning, which introduce spatial discontinuities, severely affecting the effectiveness of segmentation at the pixel level, especially in dense segmentation tasks.

In this article, we propose DyMamba, a Mamba-based model featuring a dynamic scanning strategy that adaptively plans scanning paths based on local features and complexity. In addition, to address the challenges of detail prediction and small object detection, we introduce a local aware module that performs pixel-level regional processing on images. DyMamba achieves robust segmentation across diverse microscopy image types, including cell-, organelle- and tissue-scale images. Experiments on six datasets and multiple scanning strategies demonstrate the excellent performance of our method in segmenting microscopy images, achieving an average improvement of 6.9% in mDice and 4.3% in mIoU over state-of-the-art methods across all datasets.

The code is released at https://github.com/cbqBit/dymamba.